first time trying my hand at preference optimization RL... this one was trained at 25 samples (i tried to get more, it's painful to pick by hand). should have slightly improved anatomy, text and aesthetics. will implement proper DPO and make a new version later, for now this is the best i could do.
works with flash as well

update: turns out the quality of this lora is limited by the dataset, as of now. Better loss didn't produce better results. So, unfortunately i have to do a bunch more work.